Dartmouth University Geisel School of Medicine

Empirically validated psychosocial treatments for mental health conditions are resource-intensive, limited in scalability and accessibility, and fewer than half of people with a mental health disorder receive care. Rule-based and non-generative AI chatbots have lacked the capacity for therapeutic alliance, empathy, and personalized interaction. No randomized controlled trial had previously evaluated a generative AI chatbot as a clinical mental health intervention.

Dartmouth clinicians developed Therabot, a generative AI mental health chatbot, and conducted a 210-participant randomized controlled trial. Intervention arm participants (n=106) received daily prompts to engage with Therabot for 30 days, targeting major depressive disorder, generalized anxiety disorder, and eating disorders. Therabot provided targeted interventions, question prompting, empathetic responses, and supportive affirmations, with clinician monitoring for safety and appropriateness.

Users with major depressive disorder experienced a 51% reduction in symptoms at 4- and 8-week marks. Generalized anxiety disorder participants saw a 31% decrease, and eating disorder users saw a 19% decrease in symptoms. Adherence was strong, with participants using the app during an average of 24 days per month, engaging for 6.18 hours total, and sending 260 messages on average. Participants reported therapeutic alliance comparable to that of a human therapist.

• Generative AI chatbots can deliver clinically significant symptom reduction across multiple mental health conditions in a controlled trial setting.
• High engagement and therapeutic alliance scores suggest gen AI can address a key weakness of prior digital therapeutics — patient dropout.
• Clinician oversight remains necessary: 15 safety contacts and 13 chatbot corrections were required, indicating human-in-the-loop safeguards are essential for safe deployment.